Dataset

EDGAR_FILINGS_DATASET

An open, layout-preserving conversion of U.S. SEC EDGAR filings into MultiMarkdown format covers approximately 3.4 million documents submitted between January 2022 and June 2025, designed for long-context language modeling, financial reasoning, and document analysis.

Open on Hugging Face

Data preview

Subset

Loading preview...

Dataset card

EDGAR_FILINGS_DATASET is a Hugging Face release of roughly 3.4 million parsed SEC filings covering January 2022 through June 2025, distributed in Parquet under a CC BY-NC 4.0 license. The snapshot is a subset of a larger ~18.4 million-filing corpus described in an anonymous NeurIPS 2026 submission.

What this dataset contains

Each row corresponds to one SEC filing rendered as a layout-faithful Markdown document, with embedded metadata such as CIK, SIC code, form type and period of report. The train split contains 3,775,646 rows spanning HTML, XML, SGML and PDF-OCR source formats, with the has_ocr boolean flag indicating which filings were processed via OCR. A small reviewer subset is included that pulls 25 rows from each year-month in the release window.

Structure and fields

The published schema is flat and tabular. Columns include accession and file_stem (string identifiers), year (int16) and month (int8) for filtering, parsed_md (a large_string holding the reconstructed filing text), char_count (int64), md5 (string hash) and source_format (string). Each row is self-contained, so analysts can recover the canonical EDGAR URL from the accession field without joining other tables.

How analysts use it

The release is framed for text-generation, question-answering and table-question-answering tasks over corporate disclosures, supporting queries against 10-Qs, 8-Ks, Form 4 insider trades and NPORT-P fund filings. Investment-fund use cases that fit the facts include bulk extraction of risk factors or MD&A passages, longitudinal tracking of filing volume by year and month, and building retrieval pipelines over recent U.S. public-company disclosures rather than relying on ad-hoc EDGAR scraping. The non-commercial license and ~50 GB full-corpus footprint are practical constraints to weigh before integration.

Documentation & references 1